Glossary

All terms are alphabetically organized.

Anaconda

an open-source distribution of Python (and R) that comes with many useful packages for data science, including conda

Anaconda Prompt

the command line interface for conda on Windows machines

annotation

placement of the specific label(bodypart) on an image

base environment

the default Python installation that comes with Anaconda or Miniconda and includes core Python packages; a good rule of thumb is to never install new packages into the base environment to avoid corrupting it – use virtual environment instead

batch size

number of images processed in one iteration of training (max value constrained by GPU memory). More precise term is mini-batch

benchmarking

the practice of objectively comparing machine learning tools to identify the best-performing ones

bodypart

also called label, in DeepLabCut is a arbitrarily chosen part of the animal that the user wants to track

Bonsai

a visual reactive programming language that can be used to generate complex experimental workflows. Its Bonsai.Deeplabcut package uses DeepLabCut-live for real-time markerless pose estimation.

Colab

Google Colab is a web based notebook used for writing and executing python code

conda

an open-source system for managing packages and environments; comes with Anaconda and Miniconda

CPU

central processing unit, also known as the processor, - a key component of any computing device

cropping

cutting out part of the image, used for reducing computational expense

CUDA

a parallel computing platform developed by NVIDIA for utilisation of NVIDIA GPUs

environment

see virtual environment and base environment

dataset

collection of annotations linked to specific images

detection

placement of the label by the model

GPU

graphics processing unit, also known as a “graphics card”, capable of high-throughput parallel processing; it greatly accelerates deep learning compared to a CPU

ground truth

coordinates of the label in the immage annotated by the user

identity

in multianimal DLC a parameter used for specyfing that annotations are subject specific (user can tell the difference between individuals)

IID

Independent and Identically Distributed, term used for variables that have same probability distribution but are independent from each other (a coin toss always has a 50% chance to be heads or tails, no matter what the previous result of a coin toss was)

inference

applying a trained model on data i.e. analysis

iteration

one pass of the batch through the network

jitter

natural tendency of inferred data to slightly move between frames of analyzed video. Stems from inference happening on image by image basis

MAE

mean average Euclidean error – a metric that quantifies the Euclidean distance (what we intuitively understand by the word ‘distance’) between two observations, such as the manually added and predicted bodypart labels in DLC; proportional to RMSE

Markdown

a lightweight markup language for creating formatted text using a plain-text editor

Miniconda

a lightweight version of Anaconda that includes only conda and Python, albeit with fewer pre-installed packages; useful if storage space is of concern, but greater familiarity with the command line might be required

MoSeq

developed in Datta’s Lab, an unsupervised machine learning method used to parse mouse behavior

OOD

Out-of-Domain, a term used to define data that was not used in training the model (a different dataset)

outlier

a frame in which model made bad detections

package

a specifically organised collection of Python modules (simply put: Python code) that achieve a common goal; examples include DeepLabCut and TensorFlow

path

a string of characters that uniquely defines a file or folder location on a computer, e.g. C:\Users\username\Downloads

project

the folder structure and all its contents made during project creation and later work

refinement

step of the workflow used for correction of bad detections on a subset of outlier frames

RMSE

root-mean-square error, measure of difference between values predicted by the model and ground truth

SimBA

developed in Golden’s Lab, a framework for training a supervised behavior annotation model

shuffle

in DeepLabCut: a particular instantiation of train and test sets; multiple shuffles are used for model benchmarking

snapshot

current state of the model with specific weights learned during training

supervised

a model trained using labeled data with the goal of predicting the labels on unseen data.

terminal

the command line interface for conda on MacOS/Linux machines

training

process in which the model is learning to find weights that will allow it to solve assigned task (tracking bodyparts)

tracklet

a fragment of a trajectory, relevant in multi-animal tracking. Tracklets are represented as nodes of a graph, whose edges encode the likelihood that a connected pair of tracklets belongs to the same trajectory

unsupervised

a model trained without using human annotation, only patterns from the data

VAME

developed by Kevin Luxem and Pavol Bauer, a framework for unsupervised behavior clustering

virtual environment

a self-contained Python installation that lets users have different versions of the same Python packages on a single machine; great for project management and reproducibility

when you open the terminal (MacOS/Linux) or Anaconda Prompt (Windows), the environment you are currently in is displayed in brackets, e.g. (env) user@MacBook-Pro ~ % or (env) C:\WINDOWS\system32>

weights

parameters of a neural network used to process the input (images for DLC)