Category Archives: AI Copyrights

Neural Network Applications to Image Copyright Infringement

Neural network image copyright applications

As neural networks and artificial intelligence expand, companies such as Google (GOOG: Alphabet) can employ neural networks for the purpose detecting copyright infringement in photos used on websites (blogs) and uploaded to the internet.  Neural network image copyright applications will no doubt be an application of artificial intelligence in the intellectual property arena.  Having the ability to scan hundreds of thousands of images a second, an AI neural network can be tasked to look for a certain image. Computer imaging technology can convert a simple picture into a digital matrix (e.g., a “red, green, blue” matrix) of digital values which is fed into the neural network.  From there images across the internet can be searched using Logistic Regression to determine which images likely infringe the copyrighted image.

How to understand a neural network search for copyrighted images

To understand how a neural network image copyright search would apply to the real world, we can think of the neural network as a dog which is tasked to “sniff” out the smell of a lost child. The child’s personal item (e.g., a jacket belonging to that lost child) is digitized (akin to the digitization of a copyrighted image) and is placed under the nose of the dog (fed into the neural network) so that it can obtain the scent of the child. The dog then proceeds to search for the lost child (just as the neural network searches for images that are identical to [or “substantially similar” to, as copyright law would have it]), thus finding the lost child.

Using Logistic Regression to Explain How a Neural Network Would Search For a Copyrighted Image

Techy Stuff: To get a bit into the mechanics of how the ocean of images on the internet could be classified as being a “match” or not to the image fed into the neural network, the computer would “score” the probability of whether each image is a match or not. As such, the neural network would employ something called “Logistic Regression,” where it identifies each of the data points of the digitized image as “x”, and it would plot the values on a y = wx + b graph. However, since we are looking for probability values between zero (0) and one (1), we would be looking for the “sigma” of wx + b / z [where “z” is the x-axis, and “sigma” values range from almost 0 (0% match) to almost 1 (100% match), although never touching 0% or 100% because we can never be 100% sure of whether an image is a match or not]. The output of this (sigma ((w x + b) / z)) is called “y hat”.

With artificial intelligence and the deployment of neural networks, an intellectual property right holder (a copyright holder) could easily search for and identify images that are substantially similar to their copyrighted image.

Techy Stuff: With the application of logistic regression, the way a neural network “understands” whether the results it is outputting [based on the algorithm(s) you fed it] are accurate or not. A neural network does this by comparing the “y hat” to the “y” (a digitized version of the original item). This comparison between the “y hat” output and the original is called the “loss function” for that particular output. This is akin to comparing a random object the dog searching for the lost child comes across as matching the scent of the lost child. In the copyright infringement analogy, the loss function is the comparison of the image posted on a website to the copyrighted image).

Measuring and adjusting the efficiency of neural networks using loss functions and cost functions.

Cost functions are something slightly different. A cost function measures the overall effectiveness of a search, taking into consideration the outputs of the loss function. Once the cost function (“J”) is obtained (to learn how effective the outputs of a neural network is), a neural network would then begin to employ something called a “gradient descent.” Think of the results of the cost function as an inverted parachute, and on every point of the inverted parachute (except the center at the lowest point), there is a slope. The lowest point is the point where the results of the neural network are most accurate. Thus, the neural network will analyze the slope at every point and it will sidestep either to the right or the left (on the parachute) until it finds itself at the lowest point. This allows the neural network to minimize its cost function and output the most accurate results.

[NOTE: While I am attempting to simplify the idea of the concepts of “logistic regression,” a “loss function” and a “cost function,” the point is that in order to understand what a neural network is, it is important to understand what a neural network does. Thus far, it appears to me as if a neural network employs a logistic regression in order to compare one object to another.]

How neural network image copyright applications can and will be misused

While the difficulties which arise as the copyright holder then retains an attorney to seek significant monetary damages are a direct result and will be a significant issue to contend with in coming years and decades, the topic of monetary enforcement of the copyright rights are outside the scope of this article.

Bottom Line about neural network image copyright applications

In sum, we already know that a computer can digitize an image by separating out the image into sections (a matrix), and it can assign values to the red, green, and blue matrices. These matrices can be fed into a neural network just as the jacket of a lost child can be given to a dog to obtain the child’s scent. Through logistic regression, the neural network can scan and rate the probability of whether a particular image is a match (“y hat”) for the image being searched for. Further, the neural network can identify locations and websites on the internet where the image has been used (from there, it could be fed into a simple database to 1) identify the owner of the website, and 2) to match those owners displaying the copyrighted image with the copyright holder’s own database of individuals and companies (licensees) who have paid for the right (who have obtained a paid license) to display that copyrighted image on their website.

Neural Networks, Deep Learning, and AI applications to Patent Law

Neural Networks and Artificial Intelligence Advances will need to benefit from the Intellectual Property Laws

Neural networks and real world applications of artificial intelligence will require the protection of the intellectual property law in order to benefit the companies and programmers creating the new innovations. In the coming weeks and months I will be attempting to apply concepts of deep learning and neural networks to copyright law and patent law. Most of you know my name is Robert Z. Cashman, and I am a patent attorney and the owner of the Cashman Law Firm, PLLC located in Houston, Texas. While most of my time is spend defending clients who are accused of misusing bittorrent networks, I did start my law firm with the intention of protecting ideas and furthering technology. And, in today’s fast paced world, what better technology is there to discuss other than artificial intelligence (A.I.), neural networks, and deep learning.

Are Artificial Intelligence applications of neural networks patentable?

For the budding attorney, the best way to understand neural networks is that it is code, likely copyrightable and patentable (if it accomplishes a useful goal). Obviously you can’t patent an idea, nor can you patent an algorithm (see the Metabolite v. Labcorp paper I wrote almost twelve years ago), nor can you patent the correlation between two sets of data and the interpretation thereof (a “thinking step” linking the two together). Thus, the mechanics of a neural network and how it works might not be patentable, but how that neural network is applied in the context of creating a useful result IS likely patentable (especially if it is tied to a machine).

Thus, as you can see, a neural network and the way it is programmed and applied to achieve an end result IS protectable, IS likely patentable, and thus can be understood as being PROPERTY. Thus, it can be protected with a patent, it can be sold or assigned to another individual or entity, and it can be copied or stolen in violation of the copyright laws.

I am getting ahead of myself. Let’s start off with some basic definitions so that any non-scientific person will understand how and why a neural network or deep learning in general could be useful to them.

What is a neural network and how can artificial intelligence make use of AI to provide a useful result?

A neural network is a program that uses data fed into it in order to output a result.

For example, as explained by Andrew Ng, the founder of deeplearning.ai, if someone has a set of data (e.g., a list of homes in a particular zip code, along with various houses that have been sold and for what prices), and they also have other sets of data (such as how many bedrooms a family of a particular size requires, and the walking distance of each house in a neighborhood to the local school), a neural network can crunch the data to determine which house is most appropriate for which family (e.g., which house has the correct number of bedrooms), which house best meets that family’s needs based on the ages of their children (e.g., which house is closest to the school), and the predicted price of that house based on past sales, an artifical intelligence neural network would be able to help that family choose the best home for the least amount of money that satisfies that family’s needs.

Size of the neural network matters.

In the artifical intelligence world, the size and type of the neural network required to achieve a particular result will become a relevant consideration. For example, if a local real estate broker is looking to provide an artificial intelligence service that can help place families into the best homes for their needs, and he is only looking to do so in a small area (e.g., a small dataset), he would need the use of only a small neural network.

However, if that real estate broker was looking to scale up and expand the scope of that service to the entire state, or even all of the real estate in the US, he would require a significantly larger neural network.

SUMMARY: AI applications of neural networks are likely copyrightable and patentable.

In sum, the takeaway from this article is that artificial intelligence runs its code using neural networks, and the code itself is likely copyrightable, and the application of that code is patentable. The code itself is protected from others stealing, copying, or using the code without authorization, and doing so can be either considered copyright infringement or patent infringement, depending on how the neural networks are coded.


[CONTACT AN ATTORNEY: If you have a question for an attorney about neural networks, deep learning, or artificial intelligence applications in general (or how to protect something you have created), you can e-mail us at info[at]cashmanlawfirm.com, you can set up a free and confidential phone consultation to speak to us about your AI application, or you can SMS or call us at 713-364-3476 (this is our Cashman Law Firm, PLLC’s number)].

CONTACT FORM: If you have a question or comment about what I have written, and you want to keep it *for my eyes only*, please feel free to use the form below. The information you post will be e-mailed to me, and I will be happy to respond.

NOTE: No attorney client relationship is established by sending this form, and while the attorney-client privilege (which keeps everything that you share confidential and private) attaches immediately when you contact me, I do not become your attorney until we sign a contract together.  That being said, please do not state anything “incriminating” about your case when using this form, or more practically, in any e-mail.