Artificial Intelligence Generation of Advertisements
In a generative adversarial neural network (GAN) system, a trained automatic ad generator generates ads based on product info, consumer profile, and presentation context. Ads can be vectors. Ad presentations generate metadata vectors. An effective ad discriminator is trained on ad a presentation vectors labeled by the results of ad presentations. Large numbers of ads, never presented, are generated and labeled by the effective ad discriminator and a product info decoder as positive or/and negative training example corpora. The training corpora retrain or train a new ad generator.
1 . A system for applying artificial intelligence to optimize the effectiveness of ads, the system comprising:
an effective ad discriminator that consumes generated ads and labels them as to whether they are effective, the effective ad discriminator being trained on engagement data labeled by effectiveness;
a product info decoder that consumes generated ads and decodes product info in the ads; and
an ad generator that consumes product info and generates ads, the ad generator being trained on labeled example ads for which decoded product info corresponds to input product info.
2 . The system of claim 1 wherein the engagement data is previously generated ads and engagement success is measured by consumer engagement with each previously generated ad.
3 . The system of claim 1 wherein the ads are visual images.
4 . The system of claim 1 wherein the engagement is by clicking.
5 . The system of claim 1 wherein the ads are generated audio.
6 . The system of claim 1 wherein the engagement is by a natural language expression interpreted as expressing interest.
7 . An ad generating adversarial neural network system comprising:
a conditional ad generator neural network that generates ads conditioned on source ad product info and consumer profile info;
a product info decoder function that consumes the ads and decodes the ad product info; and
an effective ad discriminator neural network, trained on ads labeled as effective, that consumes generated ads and produces a prediction of their effectiveness,
wherein the ad generator is trained on ads for which decoded ad product info corresponds to source ad product info, the training using backpropagation of probability estimates from the effective ad discriminator.
8 . The system of claim 7 wherein the ads are visual images.
9 . The system of claim 7 wherein engagement is by clicking.
10 . The system of claim 7 wherein the ads are generated audio.
11 . The system of claim 7 wherein engagement is by a natural language expression interpreted as expressing interest.
12 . A method of applying artificial intelligence to optimize the effectiveness of ads, the method comprising:
generating ads from product info input using an ad generator;
presenting the ads to consumers;
measuring the effectiveness of the presented ads; and
labeling the presented ads according to their effectiveness,
wherein the ad generator is trained on previously generated ads labeled as positive examples by an effective ad discriminator, the effective ad discriminator having been trained on prior labeled presented ads.
13 . The method of claim 12 wherein the ad generator constrains the generated ads according to consumer profile input.
14 . The method of claim 12 wherein the labeling of positive examples is conditional upon success decoding product info from the previously generated ads.
15 . The method of claim 12 wherein the ads are visual images.
16 . The method of claim 12 wherein the ads are audio.
17 . The method of claim 12 wherein the effectiveness of the presented ads is measured by consumer engagement.
18 . The method of claim 17 wherein the engagement is by clicking.
19 . The method of claim 17 wherein the engagement is by a natural language expression interpreted as expressing interest.