Task Generators

This document explains the task generation philosophy and available paradigms in the CANNs library.

Overview

The task module (canns.task) generates experimental data for CANN simulations with support for saving, loading, importing, and visualization. It provides standardized paradigms that abstract common experimental scenarios, ensuring reproducibility and convenience.

📊 Generate Input Sequences

Create time-varying external inputs that drive network dynamics

🎯 Provide Ground Truth

Supply trajectory information for analysis and comparison

Task Categories

Tasks are organized into two main categories based on the cognitive function they model.

Tracking tasks simulate scenarios where the network follows an external signal. The bump of activity in the CANN tracks a moving stimulus position.

🔵 Population Coding

Network receives static input at a fixed location. Tests basic attractor stability and population representation accuracy.

🎯 Template Matching

Network receives brief, possibly noisy inputs. Tests pattern completion and recognition capabilities.

🔄 Smooth Tracking

Most common paradigm

Network receives continuously moving input signals. Tests dynamic tracking ability with varying speeds and directions.

Available implementations:

  • SmoothTracking1D: One-dimensional tracking for ring networks

  • SmoothTracking2D: Two-dimensional tracking for torus networks (under development)

Navigation tasks simulate spatial movement scenarios [4] where the network receives velocity or heading information rather than direct position inputs through path integration [27, 28].

🔁 Closed-Loop Navigation

Network updates its internal representation based on self-motion signals. Feedback from the environment can correct errors.

➡️ Open-Loop Navigation

Network integrates velocity inputs without external feedback. Tests path integration capabilities and accumulation of errors over time.

Note

Navigation tasks do not require direct model coupling because they provide richer data (velocity, angles, etc.) that users interpret based on their specific application.

Model-Task Coupling

Why Coupling Exists

Tracking tasks require a CANN model instance to be passed during construction:

task = SmoothTracking1D(cann_instance=cann, ...)

Important

This coupling exists because tracking tasks need access to cann.get_stimulus_by_pos(). This method converts abstract position coordinates into concrete neural input patterns that match the network’s encoding scheme.

The coupling provides user convenience:

  • Automatic stimulus generation matching network topology

  • Consistent encoding between task and model

  • Reduced boilerplate for common use cases

When Coupling Is Required

Task Type

Requires cann_instance

Data Provided

Tracking tasks

✅ Yes

Input patterns in neural space
Population Coding
Template Matching
Smooth Tracking

✅ Yes

Use model.get_stimulus_by_pos() internally

Navigation tasks

❌ No

Velocity, heading, position data
Closed-Loop Navigation
Open-Loop Navigation

❌ No

Users decide how to convert to neural inputs

Design Rationale

This distinction reflects the different nature of these paradigms. Tracking involves direct sensory input to the network, while navigation involves internal state updates based on self-motion.

Task Components

⚙️ Task Configuration

Tasks are configured through constructor parameters:

  • Target positions: Where the stimulus appears or moves to

  • Durations: How long each segment lasts

  • Time step: Temporal resolution (from bm.get_dt())

  • Additional parameters: Speed profiles, noise levels, initial conditions

📊 Data Generation

The get_data() method returns:

  • Input sequence: Array of neural inputs over time (for tracking tasks)

  • Trajectory information: Position, velocity, time stamps

  • Metadata: Task parameters for documentation

💾 Data Persistence

Tasks support saving and loading:

  • save(filename): Store task data for reproducibility

  • load(filename): Reload previously generated tasks

  • Standard formats ensure compatibility

📥 Trajectory Import

Feature under development

The library supports importing external trajectories from experimental recordings. This enables:

  • Replay of real animal movement paths

  • Validation against experimental data

  • Comparison of model predictions with neural recordings

Task Usage Patterns

Standard Workflow

Typical Usage Steps

  1. Create model instance

  2. Configure task with positions and durations

  3. Generate data using get_data()

  4. Run simulation feeding task inputs to model

  5. Analyze results comparing model output to task trajectory

Multiple Trial Generation

Tasks support generating multiple trials with:

  • Same paradigm, different random seeds

  • Systematic parameter variations

  • Batch processing capabilities

Parameter Sweeps

Combine tasks with analysis pipelines to:

  • Test model robustness across conditions

  • Find optimal parameter ranges

  • Characterize attractor properties

Design Considerations

⏱️ Time Step Consistency

Tasks use bm.get_dt() to ensure temporal resolution matches the simulation environment.

Always set the global time step before creating tasks:

bm.set_dt(0.1)
task = SmoothTracking1D(...)
🎯 Position Encoding

Tasks operate in abstract feature space (angles, coordinates). The conversion to neural activity patterns is handled by:

  • model.get_stimulus_by_pos() for direct coupling

  • User-defined encoding for decoupled scenarios

🔧 Extensibility

Custom tasks can be created by:

  • Inheriting from base task classes

  • Implementing required data generation methods

  • Following conventions for output formats

Summary

The task module provides:

1️⃣

Tracking Tasks: Direct stimulus following (population coding, template matching, smooth tracking)

2️⃣

Navigation Tasks: Self-motion integration (closed-loop, open-loop navigation)

3️⃣

Model Coupling: Automatic stimulus generation for tracking tasks

4️⃣

Flexibility: Navigation tasks allow user-defined input interpretation

Tasks abstract experimental paradigms into reusable components—enabling systematic study of CANN dynamics across standardized conditions. The coupling design balances convenience for common cases with flexibility for specialized applications.