


My neural network is trained (from scratch) to get further away from the goal
This is my first time creating a neural network and I decided to create it in golang which is usually not the language used for this purpose , but I want to get a good understanding of how they work from scratch with only the basic libraries.
The goal of this program is to train a neural network to be able to add two numbers (1-10). To do this, I created a neural network class called rawai (the best name I could think of) and gave it a 1 input layer (array of size 2), 1 hidden layer (array of size 2 array) and 1 output layer (array of size 1).
The weight has two 2d arrays, one is ih (hidden input) [2,2], and the other is ho, [2,1].
The following is the code to start ai, train and test ai. You will see several debugging statements that I have used and any other functions that are not golang or its packages will be shown in the following code of my rawai class. This is called by my main function:
func additionneuralnetworktest() { nn := newrawai(2, 2, 1, 1/math.pow(10, 15)) fmt.printf("weights ih before: %v\n\nweights ho after: %v\n", nn.weightsih, nn.weightsho) //train neural network // for epoch := 0; epoch < 10000000; epoch++ { for i := 0; i <= 10; i++ { for j := 0; j <= 10; j++ { inputs := make([]float64, 2) targets := make([]float64, 1) inputs[0] = float64(i) inputs[1] = float64(j) targets[0] = float64(i) + float64(j) nn.train(inputs, targets) if epoch%20000 == 0 && i == 5 && j == 5 { fmt.printf("[training] [epoch %d] %f + %f = %f targets[%f]\n", epoch, inputs[0], inputs[1], nn.outputlayer[0], targets[0]) } } } } // test neural network a := rand.intn(10) + 1 b := rand.intn(10) + 1 inputs := make([]float64, 2) inputs[0] = float64(a) inputs[1] = float64(b) prediction := nn.feedforward(inputs)[0] fmt.printf("%d + %d = %f\n", a, b, prediction) fmt.printf("weights ih: %v\n\nweights ho: %v\n", nn.weightsih, nn.weightsho) }
The following is all the code in the rawai file:
type RawAI struct { InputLayer []float64 `json:"input_layer"` HiddenLayer []float64 `json:"hidden_layer"` OutputLayer []float64 `json:"output_layer"` WeightsIH [][]float64 `json:"weights_ih"` WeightsHO [][]float64 `json:"weights_ho"` LearningRate float64 `json:"learning_rate"` } func NewRawAI(inputSize, hiddenSize, outputSize int, learningRate float64) *RawAI { nn := RawAI{ InputLayer: make([]float64, inputSize), HiddenLayer: make([]float64, hiddenSize), OutputLayer: make([]float64, outputSize), WeightsIH: randomMatrix(inputSize, hiddenSize), WeightsHO: randomMatrix(hiddenSize, outputSize), LearningRate: learningRate, } return &nn } func (nn *RawAI) FeedForward(inputs []float64) []float64 { // Set input layer for i := 0; i < len(inputs); i++ { nn.InputLayer[i] = inputs[i] } // Compute hidden layer for i := 0; i < len(nn.HiddenLayer); i++ { sum := 0.0 for j := 0; j < len(nn.InputLayer); j++ { sum += nn.InputLayer[j] * nn.WeightsIH[j][i] } nn.HiddenLayer[i] = sum if math.IsNaN(sum) { panic(fmt.Sprintf("Sum is NaN on Hidden Layer:\nInput Layer: %v\nHidden Layer: %v\nWeights IH: %v\n", nn.InputLayer, nn.HiddenLayer, nn.WeightsIH)) } } // Compute output layer for k := 0; k < len(nn.OutputLayer); k++ { sum := 0.0 for j := 0; j < len(nn.HiddenLayer); j++ { sum += nn.HiddenLayer[j] * nn.WeightsHO[j][k] } nn.OutputLayer[k] = sum if math.IsNaN(sum) { panic(fmt.Sprintf("Sum is NaN on Output Layer:\n Model: %v\n", nn)) } } return nn.OutputLayer } func (nn *RawAI) Train(inputs []float64, targets []float64) { nn.FeedForward(inputs) // Compute output layer error outputErrors := make([]float64, len(targets)) for k := 0; k < len(targets); k++ { outputErrors[k] = targets[k] - nn.OutputLayer[k] } // Compute hidden layer error hiddenErrors := make([]float64, len(nn.HiddenLayer)) for j := 0; j < len(nn.HiddenLayer); j++ { errorSum := 0.0 for k := 0; k < len(nn.OutputLayer); k++ { errorSum += outputErrors[k] * nn.WeightsHO[j][k] } hiddenErrors[j] = errorSum * sigmoidDerivative(nn.HiddenLayer[j]) if math.IsInf(math.Abs(hiddenErrors[j]), 1) { //Find out why fmt.Printf("Hidden Error is Infinite:\nTargets:%v\nOutputLayer:%v\n\n", targets, nn.OutputLayer) } } // Update weights for j := 0; j < len(nn.HiddenLayer); j++ { for k := 0; k < len(nn.OutputLayer); k++ { delta := nn.LearningRate * outputErrors[k] * nn.HiddenLayer[j] nn.WeightsHO[j][k] += delta } } for i := 0; i < len(nn.InputLayer); i++ { for j := 0; j < len(nn.HiddenLayer); j++ { delta := nn.LearningRate * hiddenErrors[j] * nn.InputLayer[i] nn.WeightsIH[i][j] += delta if math.IsNaN(delta) { fmt.Print(fmt.Sprintf("Delta is NaN.\n Learning Rate: %f\nHidden Errors: %f\nInput: %f\n", nn.LearningRate, hiddenErrors[j], nn.InputLayer[i])) } if math.IsNaN(nn.WeightsIH[i][j]) { fmt.Print(fmt.Sprintf("Delta is NaN.\n Learning Rate: %f\nHidden Errors: %f\nInput: %f\n", nn.LearningRate, hiddenErrors[j], nn.InputLayer[i])) } } } } func (nn *RawAI) ExportWeights(filename string) error { weightsJson, err := json.Marshal(nn) if err != nil { return err } err = ioutil.WriteFile(filename, weightsJson, 0644) if err != nil { return err } return nil } func (nn *RawAI) ImportWeights(filename string) error { weightsJson, err := ioutil.ReadFile(filename) if err != nil { return err } err = json.Unmarshal(weightsJson, nn) if err != nil { return err } return nil } //RawAI Tools: func randomMatrix(rows, cols int) [][]float64 { matrix := make([][]float64, rows) for i := 0; i < rows; i++ { matrix[i] = make([]float64, cols) for j := 0; j < cols; j++ { matrix[i][j] = 1.0 } } return matrix } func sigmoid(x float64) float64 { return 1.0 / (1.0 + exp(-x)) } func sigmoidDerivative(x float64) float64 { return x * (1.0 - x) } func exp(x float64) float64 { return 1.0 + x + (x*x)/2.0 + (x*x*x)/6.0 + (x*x*x*x)/24.0 }
The output example is as follows: As you can see, it slowly moves away from the target and continues to do so. After asking, googling and searching this site I couldn't find where my error was so I decided to ask this question.
Correct answer
I think you are using the mean squared error
and forgetting the -
after differentiating.
So change:
outputerrors[k] = (targets[k] - nn.outputlayer[k])
To:
outputErrors[k] = -(targets[k] - nn.OutputLayer[k])
The above is the detailed content of My neural network is trained (from scratch) to get further away from the goal. For more information, please follow other related articles on the PHP Chinese website!

Hot AI Tools

Undresser.AI Undress
AI-powered app for creating realistic nude photos

AI Clothes Remover
Online AI tool for removing clothes from photos.

Undress AI Tool
Undress images for free

Clothoff.io
AI clothes remover

Video Face Swap
Swap faces in any video effortlessly with our completely free AI face swap tool!

Hot Article

Hot Tools

Notepad++7.3.1
Easy-to-use and free code editor

SublimeText3 Chinese version
Chinese version, very easy to use

Zend Studio 13.0.1
Powerful PHP integrated development environment

Dreamweaver CS6
Visual web development tools

SublimeText3 Mac version
God-level code editing software (SublimeText3)

Hot Topics



OpenSSL, as an open source library widely used in secure communications, provides encryption algorithms, keys and certificate management functions. However, there are some known security vulnerabilities in its historical version, some of which are extremely harmful. This article will focus on common vulnerabilities and response measures for OpenSSL in Debian systems. DebianOpenSSL known vulnerabilities: OpenSSL has experienced several serious vulnerabilities, such as: Heart Bleeding Vulnerability (CVE-2014-0160): This vulnerability affects OpenSSL 1.0.1 to 1.0.1f and 1.0.2 to 1.0.2 beta versions. An attacker can use this vulnerability to unauthorized read sensitive information on the server, including encryption keys, etc.

The article discusses writing unit tests in Go, covering best practices, mocking techniques, and tools for efficient test management.

The article explains how to use the pprof tool for analyzing Go performance, including enabling profiling, collecting data, and identifying common bottlenecks like CPU and memory issues.Character count: 159

The library used for floating-point number operation in Go language introduces how to ensure the accuracy is...

Queue threading problem in Go crawler Colly explores the problem of using the Colly crawler library in Go language, developers often encounter problems with threads and request queues. �...

This article introduces a variety of methods and tools to monitor PostgreSQL databases under the Debian system, helping you to fully grasp database performance monitoring. 1. Use PostgreSQL to build-in monitoring view PostgreSQL itself provides multiple views for monitoring database activities: pg_stat_activity: displays database activities in real time, including connections, queries, transactions and other information. pg_stat_replication: Monitors replication status, especially suitable for stream replication clusters. pg_stat_database: Provides database statistics, such as database size, transaction commit/rollback times and other key indicators. 2. Use log analysis tool pgBadg

Backend learning path: The exploration journey from front-end to back-end As a back-end beginner who transforms from front-end development, you already have the foundation of nodejs,...

The problem of using RedisStream to implement message queues in Go language is using Go language and Redis...
