Basic A.I. to Model Cognition
GIF references: Screenshots taken by H Muzart; showing code in iPython/Python IDE and Matlab IDE, scripted by H Muzart; using tools from Python ecosystem organisations, Google Brain, MATLAB/MathWorks.
On the reciprocity between Neuroscience and Machine Learning
A note on the mutual interrelations and integration of Neurobiology and A.I. :
{ image of ... }
Obviously these have inspired each other; but it is important to remember that these are separate fields of study. Computational Neurobiosciences is about understanding biologically-plausible physiological processes and networks of neurons in the animal nervous system. Deep Machine Learning (AGI) is mainly about using mathematical/statistical optimisation learning algorithms, based on data, to optimally solve specific engineering problems, regardless of the human brain's biophysical constraints. However, a good case is made (by Hassabis, Botvinick, et al [doi: 10.1.../j.neuron.2017.0...] [10.3.../fncom.2016.0...] [link-3]) that these are integrable in so many ways, and its important for them to be.
Basic classifier (Python/Tensorflow)
Here I used open-source code snippets (from the Python ecosystem organisations, Google Brain, etc) to make simple classifiers. These remain very basic, as I am still learning.
Reinforcement Learning (RL) MDP agent (Matlab) - exploration vs exploitation
From 20th Century psychologists (Skinner, Pavlov, etc) to computer scientists (Sutton, Barto, etc) to neuroscientists studying dopamine in RL (UCL's Dolan et al; DeepMind; etc), RL has become key to many things: anything from learning at school, learning new skills/habits, or whether or not you decide to look into a new shop you pass on the streets, or if you decide to click on a video you either already liked or a similar one or a completely novel one, to completing video games, to general-purpose policies. The question is: how to solve temporal credit assignments for long-term goals - that is, how can a future rewarding large goal be reached despite many small unpleasant hurdles.
A computational model microcosm of this is what I have done in 2019, as I worked on a simple model (see BioNeuroTech - pdf, videos), which uses Markov Decision Processes (MDP) in discrete states -- decisions are made to go from one state (a physical location or a physiological/psychological state of being (e.g. a thought)) to another, via an action (physical behaviour or cognitive shift), given all the information available at the time and space given. I modify the parameters as to see how the agent behaves given more explorative vs more explorative policies. Interesting patterns are found.
This essentially ties into my other endeavours into other sections (in machine learning, subcortical structures in classical conditioning, dopamine in artificial neural networks, etc). I also aim to combine my work in Unity 3D simulations with the kinds of works that are done elsewhere (e.g. DeepMind x Unity) (see my forked repos on github). This is of particular interest because it cuts across my interests in how artificial systems can mimic the prefrontal cortex and hippocampus in reward-based episodic memory, and the recall and predictive imagination of future scenes.
Other
Future works of interest:
TBC.
