Notes on Neuroscience and AI
“Neurodynamics of Cognition and Consciousness” preface:
The main theme os this volume is the dynamics of higher cognitive functions. The authors of this volume provide a line of arguments explaining why dynamics plays [a] central role in intelligence in biological brains and in man made artifacts[.] Researcher too often make the mistake of identifying intelligence with the projection of intelligence into a brilliant symbolic form, whereas intelligence is the unity of the path (or dynamic trajectory) leading to the observed formalisms and the intermitted appearance of the formalism itself. Intelligence can be understood only in the context of the complementary nature as it is describes by Kelso and collaborators. Neurodynamics had a peculiar property described as the “edge of stability” or “metastability.” Accordingly, the brain as a complete dynamic system is in perpetual movement from one state to another. When the brain reaches a dominant state, it does not [rest] there, rather it immediately moves on and decays into an unordered state, only to emerge a moment later to another prominent state. Freeman has identified neurophysiologic correlates of this metastable wandering along the landscape of brain dynamics in terms of spatio-temporal patterns of oscillations, sudden jumps or phase transitions of local field potentials. This book explores various aspects of the neurodynamics of metastable cognitive states. It covers a wide range of research areas relates to dynamics of cognition including experimental studies, dynamical modeling and interpretation of cognitive experiments, and theoretical approaches. Spatio-temporal structures of neural activity and synchronization are manifested as propagating phase cones or phase boundaries over the cortex. Methods to detect, identify and characterize such transient structures are described. Identification fo transients is a very hard pattern recognition problem as the simultaneous overlapping dynamical processes provide a noisy and cluttered domain Advanced techniques of dynamical logic progress from vague concepts to increasingly crisp and articulate forms, which is a promising approach to detect the required complex spatio-temporal correlates of cognitive functions. SIgnificant part of the volume is devotes to the description of various components of the actin-perception cycles and sensory processing domains, from cellular to system levels, and applications in intelligence designs
A brain basis of dynamical intelligence for AI and computational neuroscience:
To motivate a brain basis of neural computation, we present a dynamical view of intelligence from which we elaborate concepts of sparsity in network structure, temporal dynamics, and interactive learning. In particular, we suggest that temporal dynamics, as expressed through neural synchrony, nested oscillations, and flexible sequences, provide a rich computational layer for reading and updating hierarchical models distributed in long-term memory networks. (p. 1) While it is difficult to answer “What is intelligence?”, it is almost as useful to answer “What is intelligence for?”: Intelligence is for adaptive behavior. Otherwise, an organism would have been better off (as in the neuromythology surrounding the sea squirt) ingesting its brain and attaching itself to a rock. A corresponding yardstick for intelligence would be the degree to which an organism or agent controls its environment in service of continued survival. Indeed, extending this assessment to novel or unpredicted situations, along ecological dimensions, should correlate with generalized problem-solving capacity. (p. 2) This not-unusual definition of intelligence puts AI (based on disembodied and nonagentic neural nets trained on datasets lacking spatial, temporal, epistemic, mnemonic, social, and/or environmental context) at a disadvantage for purposes beyond hypercompetent regression and classification. Behavior is variable and complex, but it is also hierarchically organized through time in all animals, with humans exhibiting perhaps the deepest such hierarchies. Conceptual knowledge is similarly hierarchical and demanding of flexibility, reconfigurability, and combinatoric expressiveness (cf. the compositionality and systematicity of language). High level cognition is ordered, temporal, and dynamical in that what came before conditions the meaning of what comes after, with lifelong horizons in both directions (p. 2) Typically <1–2% of possible unit-wise connections exist within the cortico-limbic circuits of the hippocampus and neocortex. The impressive combinatorics inherent in this level of sparsity give rise to the intuitive, but perhaps wishful, notion that discovering the underlying motifs, generating functions, or connectomes of synaptic connectivity will unlock the brain’s neural coding secrets. Without such sparsity, dense connectivity either reliably relaxes into pattern completion for recurrent models viz. Hopfield nets, or universal function approximation for feedforward models viz. multi-layer perceptrons and deep learning. Brains appear to do both, but also much more. Density, as in typical artificial neural nets (ANNs), collapses the space of possible network configurations to that of size and layer architecture. Having far fewer degrees-of-freedom greatly restricts structural, and thus functional, diversity. (p. 4)
This is a big lesson. As a field, we still have not thoroughly learned it, as we are continuing to make the same kind of mistakes. To see this, and to effectively resist it, we have to understand the appeal of these mistakes. We have to learn the bitter lesson that building in how we think we think does not work in the long run. The bitter lesson is based on the historical observations that 1) AI researchers have often tried to build knowledge into their agents, 2) this always helps in the short term, and is personally satisfying to the researcher, but 3) in the long run it plateaus and even inhibits further progress, and 4) breakthrough progress eventually arrives by an opposing approach based on scaling computation by search and learning. The eventual success is tinged with bitterness, and often incompletely digested, because it is success over a favored, human-centric approach. […] The second general point to be learned from the bitter lesson is that the actual contents of minds are tremendously, irredeemably complex; we should stop trying to find simple ways to think about the contents of minds, such as simple ways to think about space, objects, multiple agents, or symmetries. All these are part of the arbitrary, intrinsically-complex, outside world. They are not what should be built in, as their complexity is endless; instead we should build in only the meta-methods that can find and capture this arbitrary complexity. Essential to these methods is that they can find good approximations, but the search for them should be by our methods, not by us. We want AI agents that can discover like we can, not which contain what we have discovered. Building in our discoveries only makes it harder to see how the discovering process can be done.
Pruning neural networks without any data by iteratively conserving synaptic flow
Recent works have identified, through an expensive sequence of training and pruning cycles, the existence of winning lottery tickets or sparse trainable subnetworks at initialization. This raises a foundational question: can we identify highly sparse trainable subnetworks at initialization, without ever training, or indeed without ever looking at the data? We provide an affirmative answer to this question through theory driven algorithm design
Parameter Prediction for Unseen Deep Architectures
By leveraging advances in graph neural networks, we propose a hypernetwork that can predict performant parameters in a single forward pass taking a fraction of a second, even on a CPU. The proposed model achieves surprisingly good performance on unseen and diverse networks. For example, it is able to predict all 24 million parameters of a ResNet-50 achieving a 60% accuracy on CIFAR-10. On ImageNet, top-5 accuracy of some of our networks approaches 50%. Our task along with the model and results can potentially lead to a new, more computationally efficient paradigm of training networks. Our model also learns a strong representation of neural architectures enabling their analysis.
AutoDIME: Automatic Design of Interesting Multi-Agent Environments
Of the intrinsic rewards considered we found value disagreement to be most consistent across tasks, leading to faster and more reliable emergence of advanced skills in Hide and Seek and the maze task. Another candidate intrinsic reward considered, value prediction error, also worked well in Hide and Seek but was susceptible to noisy-TV style distractions in stochastic environments. Policy disagreement performed well in the maze task but did not speed up learning in Hide and Seek. Our results suggest that intrinsic teacher rewards, and in particular value disagreement, are a promising approach for automating both single and multi-agent environment design.