Read Ten Minimalist Artificial Intelligence Lesson Notes 04_Computer Vision

1. Fairy bees

1.1. Megaphragma mymaripenne

1.2. A tiny bee species

1.3. The third smallest insect known to man

1.4. The brain is composed of only 7400 neurons, which is orders of magnitude smaller than the brains of large insects

1.5. There is no room in the tiny body for these neurons, so in the final stage of growth, it strips out the most important nuclei within each neuron to save space

1.6. We have not yet been able to understand how such a small number of neurons can achieve such complex perception and control

1.7. Its perceptual mechanism seems to be much simpler than that of today's artificial intelligence

1.7.1. Our artificial intelligence is often trained with brute force with the help of large amounts of data, thousands of artificial neurons, and enormous computing power

1.7.2. The little fairy bee tells us that there are many more elegant and simple ways to perceive the world in nature

1.7.3. Its capabilities are unmatched by any robot

2. Sensing

2.1. Perception is an important aspect of artificial intelligence

2.1.1. Sensors are extremely important, but they are only the first step in perception

2.1.2. Faced with a continuous flow of data into its digital brain, AI must also continuously discern meaning

2.2. Without the ability to perceive the external world, our AI can only live in the digital universe, using data to think mysteriously and obscurely, but has no connection to reality

2.2.1. The senses connect them to our world

2.2.2. The camera gives them vision

2.2.3. Microphones give them hearing

2.2.4. Pressure sensors provide the sense of touch

2.2.5. Accelerometers provide a sense of direction

2.2.6. A variety of bizarre sensors commonly used in science and engineering

2.2.7. Chemical sensors are able to detect chemicals more accurately than our nose or tongue

2.3. Unmanned Vehicles

2.3.1. Lidar (3D laser scanning) to detect surrounding objects and their positions, regardless of the intensity of the light

2.3.2. Cameras can see frequencies of light that we cannot see with the naked eye, through which artificial intelligence can see thermal radiation or radio waves

2.3.3. Sensors embedded in the vehicle's motor, as well as triangulation techniques and GPS using cell towers and Wi-Fi signals, can help AI know exactly where the car is on the planet and how fast it is moving

3. Learn to see

3.1. Early work on computer vision

3.1.1. Breaking down the image into constituent elements, similar to the way the human eye was accepted at the time

3.1.2. Many ingenious algorithms detect geometry and then segment the image into clearly recognizable areas

3.1.3. Used to estimate the distance between objects in the image of a stereo camera

3.1.4. Used to track moving objects

3.1.5. Construct a three-dimensional internal model of the scene with the help of several images taken from different angles

3.2. Cannae edge detection

3.2.1. Creator John Cannae

3.2.2. One of the most popular and commonly used methods in computer vision

3.2.3. Accurate detection

3.2.3.1. True edges should be spotted and misstatements should be avoided as much as possible

3.2.4. Precise positioning

3.2.4.1. The exact position of the edge should be found correctly

3.2.5. Correct Edge Counting

3.2.5.1. Each actual edge should be detected as one edge, not multiple edges

3.3. Using statistical methods, a class of algorithms has been created to recognize faces by a set of "average facial features" (base images, or eigenfaces).

3.4. Robots can now move around with greater confidence because AI can now recognize simple shapes and track the movement of objects

3.4.1. Poor performance in the case of insufficient lighting, or when the sensor data is not perfect

3.5. Computer vision systems to detect errors and omissions in the manufacturing process for quality control, while recycling plants use it to properly sort waste

4. Neural networks

4.1. Artificial neural networks are a well-established and highly successful type of artificial intelligence

4.2. The principle is to take a high degree of simplification of the way the biological brain works, and to obtain a model from which to run in a computer

4.3. Neural networks learn by changing the weights of connections between neurons, making some connections more important and others less important depending on the input

4.4. One of the biggest problems with training neural networks is the correct data

4.5. 前馈式

4.5.1. Neurons in each layer are only connected to the next layer and not in reverse

4.6. Backpropagation

4.6.1. The computer starts with the output neurons, and updates the weights and biases layer by layer through the reverse of the neurons in each layer, and finally minimizes the error of the output

4.6.2. Common methods for training such feedforward neural networks

4.7. Convolutional Neural Networks

4.7.1. Photoreceptors on the retina (rods or cones of the human eye) are not directly connected to individual neurons, but an entire area of neurons is connected to each photoreceptor cell

4.7.2. This is a very different way of routing neural networks compared to the fully connected layers of traditional feedforward neural networks

4.7.3. A deep learning network commonly used in computer vision

4.7.4. "Depth" is precisely because it has many layers of neurons

4.7.5. The era of big data has made it much easier to feed data to these neural networks

4.7.5.1. There are millions of examples of images of almost any kind

4.7.6. In 2012, computer vision has surpassed human vision, and they can recognize objects in images with superhuman precision

4.7.7. Convolutional deep neural networks have now become so smart that we no longer need to pre-compute the features of an image

4.7.7.1. Neural networks can do all this on their own

4.8. Capsule Neural Networks

4.8.1. Added more organism-inspired hierarchies to convolutional neural networks, making them more powerful

5. Racism

5.1. Prejudice is prevalent in our society, so it is not surprising that the same bias spreads to the field of artificial intelligence

5.1.1. Technology is a reflection of our hearts as well as our biases

5.1.2. Unfairly biased results may lead to biased identification of certain groups

5.1.3. In supervised learning, the performance of an AI can only be the performance of what we have been trained on

5.2. AI is typically trained primarily on images of light-skinned men, rather than people of other genders or skin tones

5.3. Facial recognition AI may perform well on images of light-skinned men, but it is error-prone when faced with images of dark-skinned women

5.4. AI systems from industry-leading companies IBM, Microsoft, and Amazon have all misclassified the faces of Oprah Winfrey, Michelle Obama, and Selina Williams, while facing the faces of white men are no problem at all

6. Forgery

6.1. Training bias is not the only problem with computer vision

6.2. Deepfake algorithms can seamlessly replace one person's face with another's face in a video

6.2.1. Widely used in the sex industry

6.2.2. Used to misrepresent politicians

6.2.3. Committing fraud

6.2.4. It has never been more difficult to distinguish fact from fiction

6.3. Malicious Deepfake Prohibition Act

6.3.1. In 2018

6.4. Deepfake Accountability Act

6.4.1. In 2019