lecture 6 - human-like AI

Question 1

deep learning

Answer 1

allows computational models that are composed of multiple processing layers to learn representations of data with multiple levels of abstraction

Question 2

backpropagation

Answer 2

indicates how a machine should change its internal parameters that are used to compute the representation in each layer from the representation in the previous layer

Question 3

deep convolutional network

Answer 3

• filtering in a CNN serves the purpose of extracting meaningful features from the input data
• convolution of the filters with the input enables the network to reduce dimensionality, achieve spatial invariance, and capture increasingly compex features as the network deepens

Question 4

parallels of CNN architecture with the visual system

Answer 4

1. increasing RF size going up the model hierarchy

• simple cells: respond to specific orientations of stimuli in preferred image locations, reacting strongly to bars and edges
• complex cells: integrate input from multiple simple cells, enabling broader spatial response and feature detection across the image

2. mimicks the representational geometry of the inferior temporal cortex: CNN develops a similar representational stucture to the IT, which suggest deeper layers of DNNs can approximate the human brain’s representational geometry for visual processing

Question 5

biologically inspired ML systems

Answer 5

• they approach ‘human level’ accuracy in a variety of domains, and even predict human brain activity.
• this raises the possibility that such systems represent the world like we do

Question 6

major flaws of DNNs (Heaven et al)

Answer 6

1. brittleness:
   - they are highly sensitive to changes in input data, which can cause them to make mistakes that humans would not typically make.

• changes such as added noise or rotation of a learned object
• this makes them vulnerable to attacks, as demonstrated by adversarial examples

2. shortcut learning:
   - DNNs overrely on pattern recognition, but lack deeper understanding of the world.

• they focus on incidental features, such as texture, background, or locations of objects, rather than the key attributes humans would learn.
• they have no good model of how to pick out what matters, so an irrelevant change in feature input can throw it off.
• this hampers transfer of learning

Question 7

adversarial examples

Answer 7

• minimal noise added to input data can lead to significant errors, such as misreading a stop sign as a speed limit sign
• humans would see the two images as identical, whereas DNNs see them as utterly different
• this could be dangerous as hackers could take over powerful AI systems, e.g., with agents that have adversarial policies

Question 8

hampered transfer of learning in DNNs

Answer 8

• reliance on shortcut learning hinders the ability to generalize knowledge from one task or domain to another
• we see this in the way that AIs can be made to lose games by adding one or two random pixels to the screen

Question 9

do DNNs represent the world the same way we do?

Answer 9

• no, they see but they do not perceive. this means that they do not understand the world in the way we do

1. they are brittle
2. they use shortcut learning

Question 10

learning levels of CNNs

Answer 10

• CNNs learn at superficial levels, not abstract levels
• this explains the limitation that they excel at recognizing surface-level patterns like color and texture, but struggle with understanding abstract concepts like sameness and relationships between objects

Question 11

human vs AI number of training examples

Answer 11

• DNNs need more training trials.
• this allows them to perform very specific tasks with little transfer of learning
• their performance is very much constrained by the training material (within-distribution)

Question 12

ethical issue of AI

Answer 12

• if the training data is not representative of the larger world, algorithms can produce biased outputs

Question 13

problem with lack of understanding/generalizability of models

Answer 13

• AI models - especially in healthcare - often fail to generalize beyond the data they were trained on
• this lack of generalizability points to the fact that these AI systems do not truly understand the task at hand

Question 14

use of AI in cognitive neuroscience: goal of computational neuroscience

Answer 14

to find mechanistic explanations of how the nervous system processes information to give rise to cognitive function and behavior

Question 15

use of AI in cognitive neuroscience: biological plausibility of models

Answer 15

1. DNNs abstract away too much from biological reality to be of use as models for neuroscience

• e.g., some DNNs have 150 layers, the ventral stream doesnt
• DNNs are often feed-forward. what about recurrent processing.

2. these models are built to optimize performance, rather than psychological or biological plausibility
3. they are still a black box
   - DNNs merely exchange one impenetrably complex system with another

Question 16

how important is biological plausibility at the implementation level

Answer 16

• from an engineering standpoint, the focus is on what works. the objective is to build systems that perform tasks effectively and efficiently, regardless of whether they mimic biological processes
• therefore, biological plausibility should be a guide, not a strict requirement.
• i.e., how closely AI mimics human brain function can serve as a guide in the design of systems.

Question 17

do NNs learn the same way babies do

Answer 17

• no

1. they need more training trials
2. training allows them to perform very specific tasks with little transfer of learning, so their performance is constrained by the training material.
3. they learn at superficial levels, but struggle with understanding abstract concepts

Question 18

marr’s three levels of analysis

Answer 18

• framework for understanding how systems (both biological brains and AI) process information

1. computation - why (problem):

• focuses on what problem the system is trying to solve, so why this system exists and what the goal of computation is.
• in the bird analogy this is flight: the system needs to solve the problem of how to move through the air

2. algorithm - what (rules):

• examines what rules or processes are used to solve the problem. it explores specific algorithms or steps the system follows
• in the bird analogy this is the flapping of the wings and coordinated movements that enable flight

3. implementation - how (physical):

• concerns the physical substrate or hardware that implements the system. i.e., how is this computation physically realized
• in the bird analogy this is refers to the feathers, muscles, and wings, that physically allow the bird to flap and fly

Question 19

Marr’s main idea

Answer 19

• ‘trying to understand perception by understanding neurons is like trying to understand a bird’s flight by studying only feathers. it just cannot be done.’
• emphasizes that understanding complex systems requires more than looking at the physical implementation
• understanding perception and cognition requires studying the computation (the goal) and algorithm level (the processes involved).

Question 20

top-down approach to modelling

Answer 20

• aims to capture cognitive functions at the algorithmic level
• disregards biological implementation so as to focus on decomposing the information processing underlying task performance into its algorithmic components
• higher cognitive fidelity, lower biological fidelity, as they focus on replicating behavior or cognitive tasks

Question 21

bottom-up approach to modelling

Answer 21

• aims first to capture characteristics of biological neural networks, such as action potentials and interactions among multiple compartments of single neurons
• this approach disregards cognitive function so as to focus on understanding the emergent dynamics of a small part of the brain, such as cortical columns and areas, and to reproduce biological network phenomena, such as oscillations
• higher biological fidelity, lower cognitive fidelity because they aim to replicate brain physiology rather than high-level cognitive functions

Question 22

synergy between top-down and bottom-up approach

Answer 22

• usually there is a trade-off between cognitive and biological fidelity
• this can turn into synergy when;

1. cognitive constraints help clarify biological function
2. biology inspires models that explain cognitive feats (i.e., improving cognitive models)

• this synergy aims to answer the common goal of explaining how the brain gives rise to the mind

Question 23

artificial neural networks (ANNs)

Answer 23

computational models inspired by the structure and function of biological networks to model behavioral and neural data

Question 24

neuroconnectionnism

Answer 24

• seeks to align ANNs with biological principles to create better tools for understanding how the brain processes information
• i.e., using ANNs inspired by biology to model behavioral and neural data

Question 25

criticisms of neuroconnectionism

Answer 25

• ANNs inspired by biology to model behavioral and neural data fail to account for basic cognitive functions
• this means that they do not fully capture the biological or cognitive complexity of real brains
• however, arguing about the successes and failures of current ANNs is not the right approach for evaluating their potential in neuroscience. the success of a scientific program should not be judged on individual results but on its capacity to generate new insights and understanding over time (i.e., falsifiable theories about brain computation)

Question 26

how DNNs can generate novel predictions about brains

Answer 26

• different brain regions represent different visual categories that can be represented as clusters that correspond to specific categories
• DNNs can be used for image classification where clusters correspond to specific categories (e.g., bodies, faces, etc.)
• this demonstrates how DNNs can be used to predict and simulate how the brain processes various visual categories
• this helps identify novel brain regions that may play previously unknown roles in processing visual stimuli, demonstrating how ML models can inspire novel biological predictions.

Question 27

are brains necessary to understand cognition

Answer 27

• brains can provide a blueprint for cognitive models, but exact replication of them is not necessary if AI can replicate outcomes.
• it is important to find synergy between top-down and bottom-up approaches to modelling to find a balance between good cognitive and biological fidelity so we can explain cognitive processes

Question 28

dark solutions

Answer 28

• paradigm in AI that goes beyond deep learning methods
• key idea perception involves more than just visible information (like pixels in an image), but that it builds on latent factors - referred to as ‘dark matter’ beyond the visible
• this paradigm therefore suggests that AI needs to move beyond simply recognizing what is seen (what & where) and start addressing more complex questions like ‘why’ and ‘how’ things happen, which are key components of human-like common sense
• goal: to develop AI that understands underlying, invisible elements of perception and can reason about the world in a more human-like, intuitive manner

Question 29

how can we simulate human intelligence best in artificial systems?

Answer 29

1. embedding hard-code rules in DNNs
2. more variable training in richer 3D environments
3. learning from less data
4. supplementing basic pattern-matching with reasoning abilities

• also, active inference in AI models

Question 30

DL improvement: embedding hard-code rules in DNNs

Answer 30

• this is a hybrid approach that combines deep learning using learned data with symbolic AI (explicit, predefined rules)
• this uses top-down learning to advance AI. it uses built-in pre-existing knowledge, rather than bottom-up, data-driven approaches.
• this way, more stuctured knwoledge is integrated into AI systems that can help understand and reason about the physical world more like humans do

Question 31

DL improvement: more variable training in richer 3D environments

Answer 31

• training AI systems in diverse, complex 3D environments where they can interact and explore
• allows them to learn more about actions and their consequences
• this improves their ability to understand and respond to the world

Question 32

DL improvement: learning from less data

Answer 32

• instead of overwriting what has been learned, we need to develop systems that can retain learned knowledge, while efficiently learning new information
• perceptual learning, in both humans and AI, benefits significantly from prior knowledge and is central for robust perception
• current NNs perform similar to young children, which struggle more with complex recognition tasks due to local biases

Question 33

DL improvement: supplementing basic pattern-matching with reasoning abilities

Answer 33

• give DNNs the ability to create their own algorithms for reasoning
• this moves beyond just recognizing patterns to actually reasoning through problems
• this would bring AI systems closer to human-like intelligence

Question 34

active inference in AI vs. passive AI models

Answer 34

• active inference suggests that the brain acts as a generative model, where sensory data is used for both recognition and controlling & predicting the consequences of actions
• unlike passive models, active inference involve agents interacting with their environment, constantly testing predictions against real-world experiences
• this approach is essential for creating AI systems that can understand and act in the world, rather than just processing sensory input without feedback from actions

Question 35

LLMs: transformer model

Answer 35

1. embedding layer: converts sequence of words into numeric pattern
2. attention layer: computes interactions between words to capture their relationships and context
3. traditional NN: produces an output that represents a new pattern of numbers, corresponding to the “meaning” of the input.

• the embedding layer and attention layer form the transformer block
• inside LLMs there are multiple transformer blocks

Question 36

what does a transformer block do

Answer 36

each block processes the input layers to refine the understanding of various aspects of ‘meaning’ to build a deeper contextual understanding

Question 37

final output of the transformer network

Answer 37

• a probability distribution over a large vocabulary of potential next words (around 50k words)
• the model calculates the probability of each word being the next in the sequence based on the context it has learned

Question 38

can LLMs display human like intelligence - arguments for

Answer 38

• they can reason, develop theory of mind, and show early signs of intelligence akin to the cognitive abilities of a 6 year old
• with further scaling, LLMs can become superintelligent, or even a potential threat to humanity

Question 39

can LLMs display human like intelligence - arguments against

Answer 39

• current conclusions are premature, as LLMs are fundamentally just next token predictors without true understanding
• i.e., they are cultural technologies that statistically summarize human knowledge found on the internet
• example: chatGPT makes mistakes when asked about numbers without the letter “e,” indicating it doesn’t fully understand the task but merely follows patterns in the data
• example: BLINK benchmark tests multimodal LLMs (language + vision models) on 14 visual perception tasks. Human accuracy is nearly 96%, but even the best LLMs perform only slightly better than random guessing.

Question 40

Mahowald et al.: dissociate language from thought in LLMs

Answer 40

• suggests that while LLMs can produce coherent language, their internal processes might lack genuine understanding or cognition
• language is therefore not the same as thought

Question 41

Mahowald et al.: formal vs. functional competence

Answer 41

• a formal language system in isolation (like LLMs) is useless/limited if it doesnt integrate with cognitive functions beyond language, such as perception and action
• the capacities required to use language to do things in the world are distinct from formal competence and depend crucially on non-linguistic cognition

Question 42

Mahowald et al.: LLMs

Answer 42

• although LLMs are good at formal competence, their performance on functional competence tasks remains spotty
• being good at language does not equate being good at thinking

Question 43

Mahowald et al.: human-like language in AI

Answer 43

• would require models to master both formal and functional competence
• may involve building modular systems with distinct processes for each time of competence, similar to how the human brain operates (separate mechanisms for language and cognition)
• architectural modularity: building modularity into the architecture of the system
• emergent modularity: naturally inducing modularity through the training process, both through training data and the objective function

Question 44

formal linguistic competence

Answer 44

• knowledge of linguistic rules and patterns
• i.e., getting the form of the language right to be able to produce and comprehend language
• phonology, morphology, lexical semantics, syntaxt

Question 45

functional linguistic competence

Answer 45

• understanding and using language to do things in real-world contexts
• formal reasoning, world knowledge, situation modeling, social reasoning
• LLMs struggle with this

Question 46

formal reasoning

Answer 46

• logic, math, planning
• ex: fourteen birds were sitting on a tree. three left, one joined. there are now eleven birds

Question 47

world knowledge

Answer 47

• facts, concepts, common sense
• ex: the trophy did not fit in the suitcase because the trophy was too small

Question 48

situation modeling

Answer 48

• e.g., discourse coherence, narrative structure
• ex: sally doesnt own a dog. the dog is black

Question 49

social reasoning

Answer 49

• e.g., pragmatics, theory of mind
• lu put the toy in the box and left. bo secretly moved it to the closet. lu now thinks the toy is in the closet

Question 50

better tests for LLMs

Answer 50

• current tests are designed for humans, which may not accurately test the decision-making, reasoning, or cognitive abilities of LLMs
• simulation is not the same as actual understanding or cogntive instantiation
• we need better tests to evaluate cognitive abilities of AI models

Question 51

fallacies to watch out for

Answer 51

1. narrow intelligence is on a continuum with general intelligence
2. easy things are easy, and hard things are hard
3. the lure of wishful mnemonics
4. intelligence is all in the brain

Question 52

fallacy: narrow intelligence is on a continuum with general intelligence

Answer 52

• misconception that narrow AI will eventually evolve into general AI whe scaled up
• though complexity of narrow models can be increased, the diversity remains anrrow
• so they improve in their specific task performance, but are confined to their narrow domains without transfer of knowledge to different tasks

Question 53

narrow AI

Answer 53

• application specific/task limited
• fixed domain models provided by programmers
• learns from thousands of labeled examples
• reflexive tasks with no understanding
• knowledge does not transfer to other domains or tasks
• today’s AI

Question 54

general AI

Answer 54

• performs general (human) intelligent action
• self-learns and reasoning with its operating environment
• learns from few examples and/or from unstructured data
• full range of human cognitive abilities
• leverages knowledge transfer to new domains and tasks
• possibly the future of AI

Question 55

what is intelligence

Answer 55

• has transfer learning and the ability to apply learned skills to new, unseen problems at its core
• increasing the size of models alone therefore won’t lead to AGI

Question 56

fallacy: easy things are easy and hard things are hard

Answer 56

moravec’s paradox: tasks requiring high-level reasoning are easy for computers, while tasks that seem simple for humans are difficult for machines to replicate

Question 57

fallacy: the lure if wishful mnemonics

Answer 57

• oversimplifying AI concepts by giving them wishful/convenient names
• can lead to misunderstandings or exaggerated expectations of AI capabilities, and anthropomorphizing technology

Question 58

fallacy: intelligence is all in the brain

Answer 58

• assumes that intelligence is purely a product of the brain’s raw computing power and that by simply increasing computational resources, machines will become as intelligent as humans
• this oversimplifies the complexity of human intelligence, which is not only about computational power but also involves interaction with the environment, embodied cognition, and other factors beyond just brain activity.