Machine Learning and Neural Networks in Neuroscience

A comprehensive vocabulary review of machine learning workflows, artificial neural network architectures, biological visual processing analogies, and the computational modeling of cortical pyramidal neurons.

Machine Learning

A branch of artificial intelligence where algorithms and statistical models learn from patterns in data to perform tasks without being explicitly programmed, requiring human input for data cleaning and manual feature extraction.

Overfitting

A modeling error in machine learning where a model fits the training data too closely by memorizing noise, resulting in poor generalization and inaccurate predictions on new data.

Under-fitting

A modeling error occurring when a model is overly simplistic and cannot capture the underlying patterns or trends present within the data.

Decision Tree

A machine learning model that sorts data sequentially through branching nodes based on the fewest possible separating features.

Kernel

A 1D or 2D grid of numerical values that acts as a filter in convolution operations, determining the influence of neighboring data points or pixels to extract specific features.

Convolution

An image-processing operation that passes a kernel over an image to calculate transformed pixel values based on the weighted sum of their surrounding neighbors, generating filters such as edge detection, blurring, or embossing.

Regularization

A technique (such as $L_1$ Lasso or $L_2$ Ridge) that penalizes overly complex models during training to prevent them from fitting noise and overfitting.

Early Stopping

A training strategy where model evaluation on a validation set is monitored, and training is halted as soon as performance begins to degrade to prevent memorization of training data.

Pruning

A regularization method for decision tree models that limits tree depth to keep the model from learning overly specific branch patterns.

Dropout

A deep learning regularization technique where randomly chosen neurons are ignored during training iterations to reduce codependency and prevent overfitting.

Error

The magnitude of the difference between the true or actual output and the predicted result produced by a machine learning model.

Backpropagation

An optimization algorithm that computes the gradient of the loss function with respect to each weight and propagates error backward through network layers to adjust weights and minimize overall error.

Perceptron

The simplest neural network model for binary classification, which computes a linear weighted sum of multiple binary inputs and produces a single binary output ($0$ or $1$) via an activation threshold.

Integrate-and-Fire (I&F) Model

A computational neuron model that integrates incoming input electrode current ($I_e$) and generates a binary voltage spike ($v$) once a threshold is reached, functioning analogously to a single artificial perceptron.

Multilayer Perceptron (MLP)

A feedforward neural network formed by interconnecting multiple perceptrons across an input layer, one or more hidden layers, and an output layer.

Input Layer

The initial layer in an artificial neural network where each individual node represents a distinct feature of the raw input data.

Hidden Layer

An intermediate layer of nodes within a neural network that calculates weighted sums and applies non-linear activation functions, functioning analogously to dendritic integration in a biological neuron.

Output Layer

The final layer of a neural network that produces the resulting prediction or classification, functioning analogously to the spike generator of a biological neuron.

Deep Neural Network (DNN)

An artificial neural network composed of many hidden layers (tens to hundreds) capable of automatically learning hierarchical representations directly from complex data.

Lateral Geniculate Nucleus (LGN)

A thalamic relay structure that receives deconstructed visual information from retinal ganglion cells and transmits it directly to the primary visual cortex ($V1$).

Primary Visual Cortex ($V1$)

The primary cortical region receiving visual signals, containing simple cells that detect boundaries and lines of specific orientations to begin reassembling visual images.

Inferotemporal Cortex (IT)

A higher-order visual cortex region in the ventral stream that reconstructs complex shapes into a wholistic representation of objects and faces.

Dorsal Stream

The visual pathway projecting to the parietal lobe and middle temporal area ($MT$) that processes object location, spatial relationships, and motion (the 'where/how' pathway).

Ventral Stream

The visual pathway projecting to the temporal lobe ($V4$ and $IT$) that processes form, color, and object identity (the 'what' pathway).

Hierarchical Processing

An organizational principle in biological visual systems and deep networks where sensory input is processed sequentially through successive stages, increasing in complexity from raw pixels/edges to complete objects.

Parallel Processing

The simultaneous routing of information across separate neural streams (such as the dorsal and ventral streams) to encode multiple distinct features at the same time.

Compartmental Model

A biophysical modeling approach developed in the 1980s that uses partial differential equations to treat dendritic segments as interconnected electrical circuits with resistors and capacitors.

Discrete Binary Spike Representation

The method of representing continuous neuronal membrane voltage fluctuations as a sequence of binary values ($0$ for no spike, $1$ for a spike) divided into $1\,\text{ms}$ discrete time intervals.

Layer 5 Cortical Pyramidal Cell (L5CP)

A structurally complex cortical neuron whose input/output transformations require a temporally convolutional deep neural network with 5 to 8 layers to be computationally approximated at millisecond precision.

NMDA Receptors and Dendritic Morphology

The two primary biological mechanisms identified by Beniaguev et al. (2021) that account for the high computational complexity and nonlinear input/output characteristics of cortical pyramidal neurons.