Show HN: 필요에 따라 자체 계층 구조를 구축하는 뇌에서 영감을 받은 신경망

A hierarchical temporal neural network that learns patterns from raw sequential data, builds its own neuron hierarchy on demand, and makes predictions through a voting mechanism inspired by how cortical columns reach consensus. No training epochs. No backpropagation. No labeled data. You feed it streams of events — stock prices, text characters, sensor data — and it self-organizes. Neurons form, compete, decay, and die. The ones that make good predictions survive. This is the Node.js reference implementation. A high-performance C++ core with Python and Node.js bindings is in development. The brain is a prediction machine. Every neuron exists to predict what comes next. Learning happens when predictions fail. Each frame, the brain: - Observes — receives events from input channels (prices, characters, pixels, etc.) - Activates — finds or creates neurons for the observations - Recognizes — checks if any learned patterns match the current context - Learns connections — strengthens links between co-occurring neurons - Learns from errors — when a confident prediction fails, creates a pattern to remember the context - Votes — all active neurons vote on what happens next, weighted by level and recency - Acts — executes the winning action predictions through output channels - Decays — unused connections and patterns weaken over time Hierarchy emerges from failure. When a base neuron's prediction fails, a level-1 pattern is created. When that pattern's prediction fails, a level-2 pattern is created. Abstraction isn't designed — it's earned. Voting enables consensus. There's no central controller. Every active neuron contributes its prediction, weighted by its level in the hierarchy and how recently it was activated. Higher-level patterns carry more weight because they represent more context. Patterns override connections. When a pattern activates on a parent neuron, it suppresses the parent's raw connection predictions. This is how the brain corrects itself — patterns exist specifically to fix prediction errors. Time is structural. Temporal distance is encoded directly in connections. A connection doesn't just say "A predicts B" — it says "A predicts B at distance 3" (three frames later). This makes sequences first-class citizens. Multiple channels converge. One data stream is mediocre. Many streams together is where it gets powerful — cross-modal patterns emerge naturally when multiple channels feed into the same brain. # Clone the repository git clone https://github.com/cucar/robot_brain.git cd robot_brain # Install dependencies npm install The brain learns to trade stocks from historical price and volume data. Each stock is a separate channel — the brain discovers cross-stock patterns and makes buy/sell/hold decisions optimized by reward feedback. The included 3-hour timeframe data is ready to use — no API key needed for this demo. node run-brain.js stock-test --timeframe 3H Expected output: Final Training Results (1 episodes): ============================================================ 📈 Overall Performance: Starting Capital: $15000.00 Total Net Profit: $221157.70 Average per Episode: $221157.70 Average ROI: +1474.38% Average Per-Frame ROI: +0.110098% Total Trades: 1268 Average Trades per Episode: 1268.0 💰 Net Profit & ROI by Episode: Episode 1: $221157.70 | ROI: +1474.38%, +0.110098%/frame (1268 trades) 📊 Base Level Accuracy by Episode: Episode 1: 55.83% The brain achieves 56% base-level prediction accuracy on price movements (which is expected — markets are noisy), but the reward-weighted action selection turns that into profitable trading by learning which contexts produce better outcomes. To download new data or different timeframes, you need a free Alpaca account: - Sign up at alpaca.markets (free paper trading account) - Get your API key and secret from the dashboard - Copy .env.example to.env and fill in your credentials:ALPACA_KEY_ID=your_key_here ALPACA_SECRET_KEY=your_secret_here - Download data: node stock-download.js --timeframe=3H - Process and run: node run-setup.js stock-test --timeframe 3H node run-brain.js stock-test --timeframe 3H The brain memorizes a repeating stock price sequence across 5 episodes, reaching 95%+ prediction accuracy. This demonstrates convergence on financial data — the same learning curve seen in text memorization. Before running, adjust the hyperparameters for stock memorization: In jobs/stock-test.js , use only 3 stocks: symbols: ['KGC', 'GLD', 'SPY'], In brain/memory.js , change contextLength to 3 : this.contextLength = 3; In brain/neuron.js , change the forget rates to 0.0001 : static connectionForgetRate = 0.0001; static patternForgetRate = 0.0001; In brain/brain.js , change the error correction to 0.3 : this.errorCorrectionThreshold = 0.3; Then run: node run-brain.js stock-test --timeframe 3H --episodes 5 --no-summary Expected output: 🎯 Final Training Results (5 episodes): ============================================================ 📈 Overall Performance: Starting Capi