Latest / The Practical AI Digest / Neuro-Symbolic AI: Combining Learning With Logic

Transcript

Neuro-Symbolic AI: Combining Learning With Logic · The Practical AI Digest

• 0:00
  Welcome to the Deep Dive. Today we're digging into
• 0:02
  something pretty fundamental that's happening in AI, a shift
• 0:06
  beyond the kind of AI we mostly see today, towards systems that
• 0:10
  are, well, more flexible and really importantly, systems we
• 0:14
  can actually understand. It feels like a really pivotal
• 0:17
  moment. It really is.
• 0:18
  And you know, it's funny how the term AI itself has changed.
• 0:21
  I remember not too long ago, researchers almost avoided it.
• 0:25
  They preferred machine learning, maybe deep learning.
• 0:28
  But now it's everywhere, which is great, but it can also make
• 0:33
  it a bit fuzzy what we're actually talking.
• 0:35
  About that's a really good point.
• 0:36
  And you know, to try and cut through some of that fog, IBM
• 0:39
  actually came up with a pretty helpful way to think about
• 0:42
  different levels or types of AI. O first, there's what we mostly
• 0:45
  have now, which they call narrow AI.
• 0:48
  These are the systems you know, that are amazing at very
• 0:50
  specific things. Recognizing pictures,
• 0:52
  translating languages, playing complex games like Go.
• 0:55
  Right, the specialist systems. Exactly super powerful within
• 0:59
  their lane, but that lane is well, narrow.
• 1:02
  They have built in limits. Then way out on the horizon
• 1:06
  there's general AI that's more the sci-fi idea right?
• 1:10
  The truly autonomous machine intelligence that can do
• 1:13
  anything a human can maybe even decide its own goals.
• 1:16
  The. Stuff that gets the headlines,
• 1:18
  sometimes, worryingly. Yeah, precisely.
• 1:20
  And when you ask researchers when we might get there, you
• 1:23
  usually hear something like 2050 and beyond which.
• 1:26
  Basically means we don't really know pretty.
• 1:28
  Much it's decades away, if ever, in that form.
• 1:31
  But the really interesting part, the bit we're actively working
• 1:34
  towards now, is the middle ground.
• 1:36
  That's what IBM calls broad AI. Broad AI.
• 1:40
  OK, So what does that look like? So this isn't science fiction.
• 1:42
  It's about building systems that can handle multiple tasks, work
• 1:46
  across different domains. You know, take knowledge from
• 1:48
  one area and apply it somewhere else.
• 1:50
  They need to fuse all kinds of data.
• 1:52
  Text, images, sensor reading, structured stuff, messy
• 1:55
  unstructured stuff. They need to work everywhere
• 1:58
  from massive cloud servers down to tiny devices on the edge.
• 2:02
  And this is the crucial bit for broad AI.
• 2:06
  These systems have to be explainable.
• 2:08
  We need to know why they make decisions.
• 2:10
  They need to be secure, fair, and critically.
• 2:13
  They need to work without needing mountains of data.
• 2:16
  They have to be efficient with small data.
• 2:19
  Those sound like the big hurdles.
• 2:20
  Explainable, fair, small data. They really are.
• 2:23
  Those are the major challenges we're tackling to get from
• 2:26
  narrow to broad AI. OK.
• 2:28
  That framework, narrow, broad general, that's super helpful
• 2:34
  context. It leads to the obvious
• 2:36
  question, if narrow AI is so good at its specific jobs, yeah,
• 2:40
  why can't we just, I don't know, make it bigger?
• 2:41
  What are those specific limits holding it back from becoming
• 2:44
  broad AI? That is exactly the core
• 2:46
  question and one of the most promising answers.
• 2:48
  A key method for bridging that gap is something called
• 2:50
  neurosymbolic computing. Neurosymbolic.
• 2:53
  Yeah, it's essentially about bringing together to really
• 2:56
  powerful long standing traditions in AI in a principled
• 2:59
  way. You take the amazing learning
• 3:01
  ability of neural networks. The deep learning stuff we hear
• 3:04
  so much about. Right.
• 3:05
  And you combine it with the logic, the reasoning, the kind
• 3:09
  of knowledge representation that symbolic AI has always focused
• 3:14
  on. OK, so you're saying deep
• 3:15
  learning, despite its successes, I mean, we see it generating
• 3:19
  art, captioning photos, beating world champions.
• 3:22
  It has fundamental weaknesses, cracks in the foundation, you
• 3:25
  might say. That's a good way to put it.
• 3:27
  The achievements are spectacular, no doubt.
• 3:29
  But yes, underneath there are some pretty significant
• 3:32
  limitations. One of the biggest is what we
• 3:35
  call brittleness. Right.
• 3:37
  Yeah, deep learning systems are essentially incredibly
• 3:40
  sophisticated pattern mattress. They get really good at
• 3:43
  recognizing patterns they saw over and over in their training
• 3:46
  data, but they often don't develop a real conceptual
• 3:51
  understanding of what those patterns mean.
• 3:54
  Can you give an example? Sure, there's a famous one
• 3:56
  involving art merit. Oppenheim sculpture, The teacup
• 3:59
  covered in fur. Oh, I think I know that one
• 4:01
  weird piece. Definitely memorable.
• 4:03
  So you show an image of this furry teacup to a top of the
• 4:07
  line image recognition system, and often it'll confidently
• 4:10
  label it teddy bear. A teddy bear?
• 4:12
  What? Well, think about it from the
• 4:13
  system's perspective. It sees something roundish, soft
• 4:18
  looking, furry. What does it know that fits that
• 4:21
  pattern? Teddy bears.
• 4:23
  It's seen in its training data. It doesn't understand the
• 4:26
  concept of a teacup and the separate concept of fur as a
• 4:29
  material and then reason that they've been combined in an
• 4:32
  unusual way. It just matches the surface
• 4:34
  features, and because the specific combination is weird,
• 4:37
  it's out of sample. It falls back on the closest
• 4:40
  pattern it knows. So it fails on edge cases
• 4:43
  basically. Exactly.
• 4:44
  Or things just slightly outside its experience.
• 4:46
  We see this in image captioning too.
• 4:48
  Systems might describe a photo of an airplane on a runway as a
• 4:51
  man riding a motorcycle on the beach.
• 4:53
  That's. Not even close.
• 4:55
  Right. The words might sound vaguely
• 4:57
  plausible together, but it shows a complete lack of understanding
• 5:00
  of the actual scene. It's just stringing together
• 5:02
  statistically likely labels based on pixels, not
• 5:05
  comprehending context. OK, so it's matching pixels, not
• 5:08
  meaning. And you mentioned training data,
• 5:10
  that seems key. Their success feels tied to
• 5:14
  having huge amounts of data, doesn't it?
• 5:15
  Is that another limit? Absolutely.
• 5:17
  That's the second major point. The deep learning revolution,
• 5:21
  sort of post 2012, was really fueled by two things happening
• 5:25
  together. One huge advances in computing
• 5:28
  power, especially using GPU's. Right, the graphics cards.
• 5:31
  Yeah, and two, the availability of massive labeled data sets.
• 5:36
  Imagenet is the classic example where millions of images all
• 5:41
  carefully categorized by humans. So they need that scale.
• 5:44
  They're incredibly data hungry. Often you need thousands, even
• 5:48
  millions of examples for a deep learning model to learn a task
• 5:52
  well. Which is totally different from
• 5:54
  how people learn. Totally different.
• 5:55
  Think about it. I can show you just one example
• 5:57
  of something you've never seen before, like that fur covered
• 6:00
  teacup again, maybe with its saucer and spoon.
• 6:02
  You get it instantly, you understand its properties, and
• 6:05
  then you can reason about it in new ways.
• 6:07
  Like if I showed you a broken one, you'd know it's still
• 6:09
  fundamentally the same object, just damaged.
• 6:12
  You could answer weird questions about it.
• 6:14
  A deep learning model couldn't do that from one example.
• 6:16
  Not even close. It needs that massive
• 6:18
  repetition, and this leads directly to the small data
• 6:22
  problem. Because most real world
• 6:24
  situations don't have millions of perfectly labeled pictures.
• 6:27
  Exactly. Think about applying AI in say,
• 6:31
  diagnosing rare diseases, or predicting failures and
• 6:34
  specialized industrial equipment or even complex financial
• 6:36
  modeling. You just don't have image net
• 6:39
  size data sets for those specific problems.
• 6:41
  So if AI needs big data, it can't work on small data
• 6:45
  problems. It struggles immensely, and even
• 6:47
  the big data sets like Imagenet have hidden biases.
• 6:51
  Researchers at the MITIBM lab did this interesting experiment.
• 6:55
  They took objects and put them in weird places, like a hammer
• 6:58
  on a bed, a knife in a bathroom sink.
• 7:00
  Things you don't normally see. Right, humans could still
• 7:03
  identify the objects with over 95% accuracy, but the best deep
• 7:07
  learning systems, their accuracy crashed down to like 4045%.
• 7:12
  Wow, that's a huge drop. Different context.
• 7:14
  Just from putting familiar objects in slightly unfamiliar
• 7:17
  settings, it really shows how much they rely on the specific
• 7:20
  statistical patterns they were trained on, how brittle they are
• 7:23
  outside that distribution. That fragility is striking, and
• 7:27
  I've also heard about AI systems being well fooled, vulnerable to
• 7:32
  attacks. Is that related to this
• 7:34
  brittleness? It absolutely is.
• 7:36
  That's another really concerning limitation, their vulnerability
• 7:39
  to adversarial attacks. How does that work?
• 7:42
  So someone can make tiny changes to an image's pixels, changes
• 7:45
  that are completely invisible to a human eye, but they can cause
• 7:49
  the AI to make a wildly wrong classification.
• 7:52
  Like what? There are examples where a
• 7:54
  picture of a stop sign after these tiny pixel tweaks is
• 7:57
  confidently identified by the AI as.
• 8:00
  Get this a brown teddy bear lying on top of the bed.
• 8:03
  Seriously, that's bizarre it. Is.
• 8:05
  And it's not just digital tricks.
• 8:07
  Researchers have created physical things like special
• 8:09
  T-shirts with weird patterns printed on them.
• 8:12
  If someone wears that shirt, they can become effectively
• 8:15
  invisible to AI systems designed to detect people.
• 8:18
  The person is right there, clear as day to us.
• 8:20
  Maybe the shirt looks a bit odd, but the AI just doesn't see
• 8:23
  them, even if the shirt is folded or wrinkled.
• 8:26
  That seems fundamentally weird about how they perceive the
• 8:28
  world compared to us. It really does.
• 8:30
  There's something different going on and the security
• 8:32
  implications obviously are pretty significant.
• 8:35
  No kidding. OK, so they're brittle.
• 8:37
  They're data hungry. They're easily fooled.
• 8:39
  What about transparency? Can we understand why they make
• 8:42
  a certain decision? That hits on another huge
• 8:44
  barrier, the lack of explain ability.
• 8:48
  Deep learning models are often described as black boxes.
• 8:52
  Heard that term a lot. Yeah, and it means it's
• 8:54
  incredibly difficult, sometimes impossible, to trace back why
• 8:57
  the model produced a specific output.
• 9:00
  Why did it approve this loan but deny that one?
• 9:03
  Why did it flag this image but not that one?
• 9:06
  And that's a problem for. For everything, for trust.
• 9:08
  How can you rely on a system if you don't know how it works?
• 9:11
  For debugging? If it makes a mistake, how do
• 9:13
  you fix it if you can't see the error in its logic?
• 9:16
  For deployment in sensitive areas like healthcare or
• 9:18
  finance, where regulations often require you to be able to
• 9:21
  explain decisions, it's a massive roadblock.
• 9:24
  OK. And one more, you mentioned
• 9:26
  reasoning earlier. Yeah.
• 9:27
  How do they fare with actual logic?
• 9:29
  That's probably the biggest gap for truly broad AI.
• 9:33
  They struggle profoundly with reasoning, especially
• 9:36
  compositional reasoning, understanding how parts relate
• 9:39
  to form a whole, or answering questions that require multiple
• 9:43
  logical steps. Imagine a picture with different
• 9:46
  colored blocks stacked up. You could ask a child how many
• 9:48
  blocks are to the right of the tall blue one, or what shape is
• 9:52
  the block sitting on top of the red cube.
• 9:54
  Yeah, a kid could figure that out.
• 9:56
  Right. But pure deep learning systems
• 9:59
  find these kinds of questions incredibly difficult.
• 10:01
  They might recognize block and blue and red, but understanding
• 10:05
  the spatial relationships right of on top of and combining those
• 10:09
  concepts to answer a multi step question that requires a kind of
• 10:13
  logical framework they just don't have internally.
• 10:16
  They lack that explicit grasp of relationships, cause and effect.
• 10:20
  So OK, we've got brittleness data, hunger, vulnerability,
• 10:24
  black box explanations, and poor reasoning.
• 10:26
  That's quite a list of limitations for deep learning on
• 10:29
  its own. It is, and that's why just
• 10:31
  trying to scale up current deep learning likely won't get us to
• 10:33
  that vision of broad AI. We need something more.
• 10:36
  Which brings us back to neurosymbolic computing.
• 10:39
  You said it combines deep learning with symbolic AI, the
• 10:43
  good old fashioned AI. Exactly.
• 10:45
  It sounds almost counterintuitive looking back to
• 10:48
  AI's roots, but there's a really powerful idea here.
• 10:50
  We have to go back briefly to the very beginning, that 1956
• 10:54
  Dartmouth workshop where the term artificial intelligence was
• 10:57
  actually born. OK, Right from the start, there
• 11:00
  were kind of two main schools of thought. 2 venerable traditions,
• 11:04
  you could say. What are they?
• 11:06
  On one hand, you have the precursors to neural networks.
• 11:09
  The idea was inspired by the brain focusing on learning
• 11:12
  complex patterns from data, adjusting connections weights
• 11:15
  between simple processing units. For a neural net, an apple is
• 11:19
  defined by how it looks, right? The pattern of pixels that
• 11:22
  statistically correlates with the label apple.
• 11:24
  Got it. Learning from examples.
• 11:27
  And on the other hand, you had symbolic AI, sometimes called
• 11:30
  Jew FAI, good old fashioned AI. This approach was totally
• 11:34
  different. It was about representing
• 11:36
  knowledge explicitly using symbols and logical rules.
• 11:39
  So how would symbolic AI see an apple?
• 11:42
  For symbolic AI, an apple is something specific.
• 11:45
  It's a type fruit. It gruesome apple tree.
• 11:49
  It has parts, stem, body, it's body, hash shape, roundish,
• 11:53
  haste, typical color, red OR green or yellow.
• 11:57
  You see, it's about defining concepts and relationships
• 12:00
  logically. Facts and rules, not just pixel
• 12:02
  patterns. Exactly.
• 12:03
  Now for a long time neural networks kind of languished,
• 12:06
  they hit bottlenecks, the so-called AI winter.
• 12:09
  But then, like we said, huge data and compute power arrived
• 12:12
  and they exploded. So the core idea, the thesis
• 12:15
  behind Neurosymbolic, is that maybe symbolic AI was also
• 12:18
  waiting, not necessarily for the same kind of data or compute,
• 12:21
  but waiting for neural networks to get really, really good at
• 12:23
  the perception part, the seeing and hearing.
• 12:26
  So neural Nets handle the messy real world input.
• 12:30
  Yes, and now that they can do that reliably, we can combine
• 12:33
  their strength in perception with the strength of symbolic AI
• 12:37
  in reasoning and knowledge representation.
• 12:39
  Marrying the two. Precisely.
• 12:41
  Neurosymbolic computing is about that principled integration,
• 12:45
  taking the best of both worlds. The robust learning from data of
• 12:49
  neural Nets and the logical reasoning, the explicit
• 12:51
  knowledge, the potential for explainability from symbolic AI
• 12:56
  to create something that's genuinely more powerful and more
• 12:58
  trustworthy than either approach alone.
• 13:01
  More than the sum of its parts. OK, conceptually that makes a
• 13:04
  lot of sense. But how does it actually work?
• 13:06
  Can you walk us through an example?
• 13:08
  How do you physically or well architecturally combine these
• 13:12
  two seemingly different ways of thinking about the world?
• 13:15
  Like how does the pixel pattern Apple meet the rule based Apple?
• 13:19
  Great question. Let's use a concrete example.
• 13:21
  That sea lever data set I mentioned, the one designed for
• 13:24
  complex visual reasoning. Remember questions like Are
• 13:26
  there an equal number of large things in metal spheres?
• 13:29
  Yeah, the kind pure deep learning struggles with because
• 13:32
  it mixes perception and logic. Exactly.
• 13:35
  It's hard because the concepts large metal sphere and the the
• 13:39
  reasoning compare are all tangled together in one big
• 13:43
  neural network. Humans don't do it like that,
• 13:45
  right? We first see the objects,
• 13:47
  understand the question, and then apply logic step by step.
• 13:51
  Makes sense? So a neurosymbolic system, often
• 13:53
  called something like a neurosymbolic visual question
• 13:56
  answering system, mimics that human process.
• 13:59
  It typically has a few key parts working together.
• 14:02
  First you have the vision part, which is usually a neural
• 14:05
  network like a CNN. But instead of just outputting a
• 14:08
  single label like objects, it does something more
• 14:11
  sophisticated. It essentially D renders the
• 14:13
  scene. It looks at the pixels and
• 14:15
  translates them into a structured symbolic
• 14:17
  representation. Like a description.
• 14:19
  Kind of, but very structured. Think of it like generating a
• 14:22
  database table or a scene graph on the fly.
• 14:24
  Would say object 1 shape dot cube, color dot red size large
• 14:28
  material location dot XYZ, object 2 shape sphere colors
• 14:33
  blue size, small material dot metal, location dot XYZ and so
• 14:36
  on for everything it sees. So the neural network is
• 14:39
  extracting facts from the image. That's key.
• 14:41
  That's absolutely key. It's translating from pixels to
• 14:44
  symbols. Then second you have the
• 14:46
  language part. Another neural network may be an
• 14:49
  LSTM or a transformer. Takes the natural language
• 14:51
  question like Are there an equal number of large things in metal
• 14:55
  spheres? And it translates that not into
• 14:58
  an answer, but into a symbolic program.
• 15:00
  A program. Like code.
• 15:02
  Effectively, yes, a sequence of executable steps based on the
• 15:06
  questions meaning. So that question might become a
• 15:09
  little program like scene filter size, large count, count, scene
• 15:13
  filter, material, metal filter count count 2, compare equal
• 15:16
  count 1 count 2. The network learns to convert
• 15:20
  language into logical operations.
• 15:22
  OK, so you have facts about the scene from the Vision Net and a
• 15:25
  program from the language? Net exactly and 3rd you have the
• 15:28
  reasoning part. This is often a purely symbolic
• 15:31
  component. It takes the program generated
• 15:33
  by the language NET and simply executes it on the structured
• 15:36
  scene description provided by the vision net.
• 15:39
  It runs the steps filter count compare using standard symbolic
• 15:43
  logic. And that gives the final answer.
• 15:44
  That gives the final answer, but there's a fourth crucial piece,
• 15:48
  joint training. How do they learn together?
• 15:50
  This is really clever. The neural parts, the vision net
• 15:53
  extracting symbols, and the language net generating programs
• 15:56
  aren't trained separately. First, they're trained jointly,
• 15:59
  often using reinforcement learning.
• 16:02
  The entire system gets a reward only if the final answer
• 16:05
  produced by the symbolic reasoner is correct.
• 16:08
  So the reward signal flows back through the whole chain.
• 16:10
  Precisely this forces the neural networks to learn to do their
• 16:14
  translation jobs well. The vision net learns to extract
• 16:17
  the right symbols needed for reasoning, and the language net
• 16:20
  learns to generate the right programs that lead to the
• 16:22
  correct answer. When executed, that symbolic
• 16:25
  layer, the scene description in the program, acts as a kind of
• 16:28
  meaningful choke point or interface between perception and
• 16:32
  reasoning, and that's where unlocks the benefits.
• 16:34
  That architecture is really elegant.
• 16:36
  It feels like it builds in the logic that was missing before.
• 16:40
  Okay, so beyond the clever design, what are the actual
• 16:44
  results? What are the big game changing
• 16:47
  benefits that make this approach so promising for building better
• 16:50
  AI? The benefits really are quite
• 16:53
  profound, addressing those core weaknesses we talked about.
• 16:56
  First off, it solves hard problems that were previously
• 16:59
  out of reach. On data sets like C Labar, these
• 17:03
  neurosymbolic systems achieve near perfect accuracy end to
• 17:07
  end. Deep learning models just
• 17:09
  couldn't crack it. Getting that kind of result was
• 17:11
  a big deal. It earned spotlight
• 17:12
  presentations at major AI conferences like Nuri PS.
• 17:16
  So demonstrably better performance on complex tasks.
• 17:19
  What else? 2nd, and this is maybe the most
• 17:21
  exciting part for practical applications, massive data
• 17:25
  efficiency. Remember the small data problem?
• 17:27
  Deep learning needing millions of examples?
• 17:29
  Yeah, neurosymbolic systems can achieve better performance than
• 17:32
  the end to end deep learning models using only about 1% of
• 17:35
  the training data. 1%. Seriously.
• 17:38
  Yeah, 1% and with just 10% of the data, they often reach near
• 17:41
  perfect accuracy. That's.
• 17:43
  That's revolutionary for real world AI if you only need 100th
• 17:46
  of the data. It completely changes the game.
• 17:49
  Suddenly AI becomes feasible for all those areas.
• 17:52
  Specialized manufacturing, rare medical conditions, niche
• 17:56
  financial analysis where you simply don't have massive
• 17:59
  labeled data sets. It drastically lowers the
• 18:02
  barrier to entry and broadens AI's applicability.
• 18:05
  Incredible. OK, data efficiency is huge.
• 18:08
  What about the black box problem?
• 18:10
  That's the third major benefit, explainability and debugging,
• 18:14
  because you have that symbolic program as an intermediate step
• 18:17
  that filter large count sequence, right?
• 18:19
  The generated code. You can actually look at it.
• 18:21
  When the system gives an answer, you can inspect the program it
• 18:24
  generated and executed. You can see the logical steps it
• 18:26
  took. OK, first it filtered for large
• 18:28
  objects, then it counted them. Then it filtered for metal
• 18:31
  spheres, counted those, then it compared the counts.
• 18:33
  So you can follow its reasoning. Exactly.
• 18:35
  It provides transparency. If it makes a mistake, you can
• 18:38
  often pinpoint where in the program the logic went wrong,
• 18:41
  just like debugging traditional software code.
• 18:43
  This builds trust, makes it easier to deploy, and helps meet
• 18:46
  those regulatory requirements for explainable AI we talked
• 18:50
  about. That's a massive advantage,
• 18:52
  especially in high stakes areas. Are there other benefits
• 18:55
  emerging too? Yeah, the field is moving fast.
• 18:57
  We're seeing systems capable of dynamic concept learning.
• 19:01
  So instead of needing all concepts like red or cube to be
• 19:05
  perfectly predefined, the neural part can learn to map objects
• 19:09
  into some kind of internal space, like a color space or
• 19:12
  shape space, and then learn the meaning of new concepts
• 19:15
  dynamically from how they're used in questions and answers.
• 19:18
  So it could learn what cyan means, even if it wasn't
• 19:21
  explicitly told beforehand. Potentially, yes, by seeing it
• 19:25
  used in context alongside objects it perceives.
• 19:28
  This is vital because, you know, the world's full of new concepts
• 19:31
  we need AI to adapt to. Makes sense.
• 19:34
  And taking that even further, there's work on meta concept
• 19:36
  learning. These systems try to understand
• 19:38
  relationships between concepts. For example, learning that kid
• 19:42
  is basically a synonym for child, or that car and truck
• 19:45
  both belong to the category vehicle.
• 19:47
  So it builds a kind of common sense knowledge structure.
• 19:49
  Kind of, yeah. Leveraging these relationships
• 19:52
  allows for much more flexible and powerful reasoning, making
• 19:55
  the AI more robust when encountering variations in
• 19:59
  language or needing to generalize.
• 20:01
  It's getting closer to how humans use interconnected
• 20:03
  knowledge. This visual reasoning example is
• 20:06
  really compelling, but it feels like just the start.
• 20:08
  Where else could this neurosymbolic blend be applied?
• 20:11
  What other horizons are opening up?
• 20:13
  Oh absolutely. It's much broader than just
• 20:15
  static images. People are already pushing into
• 20:18
  video reasoning. Think about understanding
• 20:20
  actions and interactions over time.
• 20:23
  Projects like Video C Lover are looking at relationships between
• 20:26
  objects and video clips, even exploring counterfactuals.
• 20:29
  What would have happened if that blue block hadn't been pushed?
• 20:32
  Understanding cause and effect in motion.
• 20:34
  Exactly. Moving beyond just description
• 20:37
  to deeper comprehension of dynamic scenes, another big area
• 20:40
  is planning. Symbolic AI has always been
• 20:43
  strong in classical planning problems.
• 20:45
  Think Tower of Hanoi, solving mazes, scheduling tasks.
• 20:49
  Neurosymbolic approaches are trying to combine that logical
• 20:52
  planning capability with the perceptual abilities of neural
• 20:55
  Nets. For instance, instead of
• 20:57
  planning purely with predefined symbols, maybe you plan within
• 21:01
  the learned latent space of a neural network that has
• 21:03
  compressed complex state information.
• 21:06
  This could open doors for more sophisticated robotics,
• 21:09
  logistics or process control. Interesting planning in a
• 21:12
  learned representation rather than a hand coded 1.
• 21:15
  Potential. Yeah.
• 21:16
  And there's one more crucial horizon we have to mention,
• 21:19
  which applies to all of AI, but where Neurosymbolic might offer
• 21:23
  some advantages. Energy efficiency.
• 21:26
  The power consumption problem. We hear a lot about how much
• 21:29
  energy these big models use. It's a massive issue training.
• 21:33
  Some of the largest models can consume electricity equivalent
• 21:36
  to powering a small city for days.
• 21:38
  It's unsustainable environmentally and economically
• 21:41
  if AI is going to scale globally.
• 21:43
  How does neurosymbolic help there?
• 21:45
  Well, partly through data efficiency.
• 21:47
  Less data often means less training compute needed.
• 21:50
  But also the research community, including places like MIT and
• 21:53
  IBM, is attacking this on multiple fronts.
• 21:56
  They're exploring entirely new types of hardware.
• 21:58
  Things like memristers, analog computing chips, maybe even
• 22:01
  quantum computing down the line, designed specifically for AI
• 22:04
  workloads and potentially orders of magnitude more efficient.
• 22:07
  Plus, new algorithms are being developed alongside these
• 22:10
  hardware ideas, all aiming to drastically slash a is energy
• 22:14
  footprint, maybe by factors of hundreds or thousands.
• 22:17
  Broad AI needs to be green AI. That seems absolutely essential
• 22:21
  for the future. So wrapping this up, it really
• 22:23
  sounds like neuro symbolic isn't just, you know, gluing 2 old
• 22:27
  ideas together. It feels like a genuinely new
• 22:29
  synthesis, creating something where the strengths of neural
• 22:32
  Nets and symbolic AI actually compensate for each other's
• 22:35
  weaknesses. That's exactly the way to think
• 22:37
  about it. It's synergistic.
• 22:39
  You get the learning power for perception and the logical
• 22:41
  structure for reasoning and explanation.
• 22:44
  And that vision is really what's driving major efforts like the
• 22:47
  MITIBM partnership, that decade long quarter billion dollar
• 22:51
  collaboration. It's about marrying MI TS deep
• 22:54
  academic research with IB Ms. Focus on real world industry
• 22:58
  problems, using real data and computational scale to push AI
• 23:02
  towards being broadly applicable, more trustworthy,
• 23:04
  and yeah, maybe a little more like us and how it understands
• 23:07
  the world, building on that whole legacy since 1956.
• 23:10
  So for you listening, what's the key take away from this deep
• 23:13
  dive into neurosymbolic computing?
• 23:15
  What does it mean practically, whether you're building AI
• 23:17
  systems or just trying to understand where this technology
• 23:19
  is headed? I think the main thing is that
• 23:21
  this isn't just theory, it's a practical path towards building
• 23:25
  AI that overcomes many of the limitations we see today.
• 23:28
  It points towards AI systems that are fundamentally more
• 23:31
  trustworthy because their reasoning can be inspected and
• 23:33
  explained. It enables AI that works with
• 23:36
  much less data, which is critical for countless real
• 23:39
  world applications currently starved for huge data sets.
• 23:42
  It leads to AI that's more robust and can reason more
• 23:45
  effectively about novel situations, moving beyond simple
• 23:48
  pattern matching. For engineers understanding this
• 23:51
  architecture, how neural Nets translate raw input into
• 23:54
  symbolic facts, how other Nets translate questions into
• 23:57
  programs, how symbolic engines execute those programs, and how
• 24:00
  reinforcement learning ties it all together, that's becoming
• 24:02
  increasingly important. It provides A blueprint for
• 24:04
  building more capable and reliable AI.
• 24:07
  It's about making AI genuinely more useful across the board.
• 24:10
  It really feels like an evolution, moving AI beyond just
• 24:14
  being incredibly good at narrow tasks, even if data hungry and
• 24:17
  opaque, towards something more versatile, more transparent, and
• 24:21
  maybe, yeah, closer to what we intuitively think of as
• 24:23
  intelligent. So here's a final thought to
• 24:26
  leave you with.